Home / AI-901

Microsoft · AI-901

AI-901 Practice Test

Free AI-901 practice test questions from all 2 official domains, with every answer option explained.

Start practicing free Take a mock examSee plans

Microsoft Azure AI Fundamentals (AI-901)

Questions
40-60

Typical range

Time
45 min
To pass
700 / 1000

Scored on a 1-1000 scale.

What is on the AI-901 exam?

Open a domain for its objectives and their practice questions.

Try free AI-901 practice questions

Pick an answer to see why each option is right or wrong. A preview of 10 of the 10 free questions, which cover every domain.

Domain 1: Identify AI concepts and capabilities

1.3 Identify AI workloadseasy

A travel company already has a generative AI chatbot that answers policy questions. It now wants the assistant to rebook flights and send confirmation emails on the customer's behalf. What does this require?

Included in Sample: sample every domain.See plans

Choose an answer to check it. Afterward, select any option to review its explanation.

Sentiment analysis tells you how a customer feels. It doesn't carry out any task.

A voice changes how replies are delivered, not what the assistant can do. Rebooking still needs actions.

Conversation history is a real way to improve a chatbot, so it looks like an upgrade. It keeps the dialog coherent, but the bot still can only reply; it can't rebook a flight or send an email.

Agents combine a language model and instructions with tools, and action tools are what let them perform tasks such as sending emails or updating systems for the user.

1.1 Describe principles of responsible AIeasy

According to the reliability and safety guidance, which kind of situation should designers plan for in addition to the conditions they expect?

Included in Sample: sample every domain.See plans

Choose an answer to check it. Afterward, select any option to review its explanation.

The guidance asks designers to plan for unintended circumstances, naming accidental interactions, malicious data and cyberattacks. Systems should respond safely and resist manipulation even when conditions are not what was expected.

Team diversity is a fairness concern about how the system is designed, not a runtime condition the system must withstand.

Dataset size and sharing are not named as conditions to design for. The guidance focuses on accidental interactions, malicious data and attacks.

Questions about training data relate to transparency and are an expected, healthy part of using AI, not an unintended circumstance that threatens safe operation.

1.2 Identify AI model components and configurationseasy

A company wants to add AI features to a lightweight mobile app that runs on the device and covers one narrow topic area. Which kind of language model suits this best?

Included in Sample: sample every domain.See plans

Choose an answer to check it. Afterward, select any option to review its explanation.

Small language models suit focused topic areas and are easy to deploy locally, for example on devices, which fits a lightweight phone app.

Image generation models create pictures from text. The app needs language features, not image output.

An LLM generalizes well across many topics, but that breadth costs more to train and use. On-device, single-topic use favors a smaller model.

Video generation is resource intensive and produces video clips. It doesn't provide language features for a phone app.

1.1 Describe principles of responsible AIeasy

Which statement describes the inclusive design principle "Solve for one, extend to many"?

Included in Sample: sample every domain.See plans

Choose an answer to check it. Afterward, select any option to review its explanation.

This Microsoft inclusive design principle says that solving for a particular need, such as one disability, often produces solutions that benefit many more people.

The inclusiveness guidance still calls for testing with people with disabilities; no single design removes the need to validate with users.

Ownership of fixes relates to accountability and remediation, not to an inclusive design principle.

Reading "extend to many" as one unchanged design for everyone misses the point: the principle starts from specific needs, not from a single generic design.

Domain 2: Implement AI solutions by using Microsoft Foundry

2.4 Implement AI solutions for information extraction by using Foundryeasy

A warehouse analyzer reads shelf photos. One field must report the stock level as exactly one of 'full', 'partial', or 'empty'. Which extraction method should that field use?

Included in Sample: sample every domain.See plans

Choose an answer to check it. Afterward, select any option to review its explanation.

Generate creates values from the image, which suits descriptions, but its free-form output wouldn't stay within three fixed levels.

Classify picks one value from a predefined list, which guarantees the field holds one of the three levels.

Extract pulls values that appear in the image, like label text; a stock level is judged from the scene, not printed on it.

Field methods are extract, classify and generate; there is no summarize method.

2.1 Implement generative AI apps and agents by using Foundryeasy

A clinic puts a deployed chat model behind its booking website. Whatever a patient types, every reply must stay friendly and must decline to give medical advice. Where should the developers place these rules?

Included in Sample: sample every domain.See plans

Choose an answer to check it. Afterward, select any option to review its explanation.

Standing rules for tone and limits belong in the system prompt: the model answers every user prompt while obeying that guidance, so the rules hold whatever patients type.

TPM governs throughput on the deployment; it cannot express rules about tone or forbidden topics.

Tempting because history is resent with each prompt, but its job is to keep earlier questions and answers in context; it records what was said rather than defining how the model must behave.

Temperature trades predictability for creativity; raising it makes replies less controlled, not rule-bound.

2.1 Implement generative AI apps and agents by using Foundryeasy

A compliance reviewer asks which setting on a Foundry model deployment decides where and how its inference requests are processed. Which deployment parameter should the team point to?

Included in Sample: sample every domain.See plans

Choose an answer to check it. Afterward, select any option to review its explanation.

Temperature is a runtime generation parameter for creativity; it says nothing about processing location.

TPM is also a deployment parameter, which is why it tempts, but it sets how much token traffic the deployment can handle, not where processing happens.

Deployment types such as standard, global batch or regional provisioned throughput determine where and how inference runs, tied to throughput and data-processing needs.

System instructions shape the model's behavior and role; they do not configure the deployment's infrastructure.

2.3 Implement AI solutions with computer vision and image-generation capabilities by using Foundryeasy

An image-generation app in Foundry receives the generated picture as base64 text. What is base64 encoding for?

Included in Sample: sample every domain.See plans

Choose an answer to check it. Afterward, select any option to review its explanation.

Base64 changes how bytes are written, not how long the model takes; it doesn't shorten generation.

The image generation model creates the picture from the prompt; base64 only packages the finished bytes.

Images are raw bytes, while JSON is text-only; base64 turns the bytes into safe text so the picture can travel inside the response, and the app decodes it back to bytes.

Base64 is an encoding anyone can reverse, not encryption; access control comes from the API key or Entra ID on the endpoint.

2.2 Implement AI solutions for text and speech by using Foundryeasy

A support tool calls detect_language() from the Azure Language SDK on a customer's message. What does the result give the app?

Included in Sample: sample every domain.See plans

Choose an answer to check it. Afterward, select any option to review its explanation.

Turning text into speech is Azure Speech's text-to-speech capability, not part of language detection.

That is what recognize_pii_entities returns; language detection identifies the language rather than masking personal data.

The method reports which language the text is in, the matching ISO 639-1 code and a confidence value from 0 to 1.

Detection only names the language; translating text is something you would ask a general-purpose model to do.

2.2 Implement AI solutions for text and speech by using Foundryeasy

A text-to-speech solution typically needs which two inputs?

Included in Sample: sample every domain.See plans

Choose an answer to check it. Afterward, select any option to review its explanation.

Audio files with SAS URIs are inputs for batch transcription, which goes from speech to text.

Confidence scores come back from analysis or recognition results; they are not what synthesis is given.

A microphone stream is input for speech recognition; synthesis starts from text, not audio.

Text-to-speech vocalizes a given text, so it needs that text and the voice that will speak it.

Practice all 10 free questionsGet 150+ questions

AI-901 question bank coverage

DomainExam weightFree sampleMaster (full bank)
1. Identify AI concepts and capabilities43%464
2. Implement AI solutions by using Microsoft Foundry57%686
Total100%10150

Looking for a AI-901 practice exam? Start with the free questions above, then take the timed mock exam to rehearse the AI-901 format under exam conditions.

How these AI-901 questions are made

The domains, weights and objectives come from Microsoft's official AI-901 exam guide (vSkills measured as of April 15, 2026). Questions are original and written only from the vendor's official prep material, and each one cites the passage it is based on. Every question is then answered independently, without seeing its key, and every answer option's explanation is checked for accuracy before it is published.